Replies to post #1331 on Dendreon Corp fka DNDNQ

[ocyanblue]

My results are empirical, not theoretical. Therefore I cannot say with absolute certainty that they are universal.

As long as the covariates considered are balanced, this is likely true in general and due to the noise filtering effect inherent in the iterative regression steps. There is quite a bit of discussion in various papers about how postieri corrections can increase the power of a study. See, for example, the section "When are adjustments made?" in the below:

http://www.tufts.edu/~gdallal/adjust.htm

[io_io]

iwfal - thnkx for answer .........


<<<"The results were pretty consistent - in 1000 trials of the same size a Cox Regression has a median improvement in p value of greater than 2x. Sometimes it was less than 2. But it never approached a 0x improvement.">>>

You probably covered these types too, but let me ask just in case - in DNDN's trials one unique thing is that there are almost no censored patients - also I wonder if your generated trials were there both "regular/classical" and "random/odd" shaped curves - but you thought of all that already ?



<<<"I suspect that they do push and generally the FDA pushes back.">>>

Just to remark that a doubling of the p value is like the difference between a 1-sided and 2-sided test - which of course the ICH guidance docs generally require, and the FDA generally gets.

Next time I see a Cox as a primary endpoint, I will be wondering what the justification is. But it does bode well for 9902b.

[io_io]

<<<"I suspect that they do push and generally the FDA pushes back.">>>



I still am puzzled by this. I have read a number of FDA stats reviews, and I have seen more than once where the alternate Cox did produce a beter p-value than log-rank.

But I have never seen the statistician reply that he/she is not the least bit surprised, given the lower hurdle. And they are at liberty to write what they want, and do tend to be descriptive. Instead they simply write about alpha, etc, and go on to comment about the co-variates nevertheless.

Odd ?